Apparently they can’t even be bothered to price shop for GPUs, as Nvidia has 94% market share at a time when Intel and AMD are selling sublime mid range cards. How can we expect folks to stop buying Call of Duty on principle?
That statistic didn’t count the bulk of ‘AI cards’ other than the rare rich dev who picks up a 5090, its almost purely gaming GPU sales. It’s gamers buying the 94%.
That 94% does not include AI GPUs at all. Just ‘pro’ cards like the RTX Pro, and gaming ones.
The vast majority of inference/training is done on top end SXM GPUs like the H200, MI325X and such, and sometimes on server only PCIe cards like the L40. But the statistic you’re quoting is for ‘Addin boards,’ which don’t include this server stuff.
There are tinkerers who run ML stuff locally on desktop GPUs, but its a pretty small market, and honestly they’re all useless for many experiments because they don’t have enough VRAM.
Point I’m making is ‘AI GPUs’ don’t skew this statistic much; it’s gamers pushing that 94% market share.
Yeah, well ‘gamers’ do what they do at scale.
Apparently they can’t even be bothered to price shop for GPUs, as Nvidia has 94% market share at a time when Intel and AMD are selling sublime mid range cards. How can we expect folks to stop buying Call of Duty on principle?
crypto ai bros and the companies they work at unfortunately
That statistic didn’t count the bulk of ‘AI cards’ other than the rare rich dev who picks up a 5090, its almost purely gaming GPU sales. It’s gamers buying the 94%.
How does the data differentiate between AI cards and gaming cards?
That 94% does not include AI GPUs at all. Just ‘pro’ cards like the RTX Pro, and gaming ones.
The vast majority of inference/training is done on top end SXM GPUs like the H200, MI325X and such, and sometimes on server only PCIe cards like the L40. But the statistic you’re quoting is for ‘Addin boards,’ which don’t include this server stuff.
There are tinkerers who run ML stuff locally on desktop GPUs, but its a pretty small market, and honestly they’re all useless for many experiments because they don’t have enough VRAM.
Point I’m making is ‘AI GPUs’ don’t skew this statistic much; it’s gamers pushing that 94% market share.
Ah ok, so specific models. Gotcha, thanks